Research On Social Recommendation And Its Privacy Preserving

With the development of Internet technologies and Web2.0 applications,online social media has become the important platforms for information diffusion,information access,information sharing and user-generated content.The popularity of online social applications makes continuously massive information diffusion which resulting in information overload.Moreover,user-generated content often contains sensitive information,which make the user under the risk of personal privacy leakage.Social recommendation is a powerful tool to deal with online information overload problem.Users' online social interaction and social attribute information are applied to social recommendation to help users to filter redundant information and to obtain a personalized recommendation service.However,how to effectively extract user social network information from online social media for social recommendation and how to help users get accurate recommendation while protecting user privacy has became the theoretical and practical research problem.In order to address the challenges of social recommendation,this thesis discusses the characteristics of social recommendation applications and summarizes privacy protection techniques.Then it takes the information extracting in online social networks and social perceptual model as the starting points for this reasearch.It studies a random walk mechanism and a privacy pretection mechansim from the perspectives of quantitative calculation of trust and hierachical portray of user social networks respectively.To study the social perceptual for social recommendation,a method to portray social context is given and a concept of multi-domain mapping,interaction,penetration is proposed,which includes social domain,cyber domain and physical domain.The online social context can be mapped into two dimensions of self-presentation and social presence.The main components of the context object can be refined from the complex online scenario to provide effective operational objects for entity information extraction.In this thesis,the social tie strength and social trust are both investigated to measure the effects on social recommendation utility.The results show as follows: it needs to consider the impact of social homogeneity on users' social attributes and the users' online social interactions and mode are proposed to form the external expression of social homogeneity based on that conclusion;Compared to the social tie strength,social trust exhibit more asymmetrical attributes and can be uesed to subdivide the features of social network entities.To build the trust-based recommendation mechanism with random walk,the following two methods are proposed.A novel trust network construction method is proposed to address the measurements,expression and construction problems on social trust.Unlike traditional binary trust method,this method utilizes social entity attributes to quantify the value of the social trust.It is a measurement on trust based on physical network structural information.The advantage is that it can integrate the cosine distance of the interaction characteristics of online entities,which can quickly discover recommendation nodes during the computational process.Experimental results show that this method can effectively discover the entities with similar online interaction characteristics.Considering the cold strart and sparse matrix problems in recommendation,a random walk mechanism based on social trust network is proposed.It uses the quantitative entity trust value as the walk preferences in the trust network and quickly finds the target nodes based on the characteristics of entities and the factor of stop probability.Experimental results on the trust network show that,the computational outcome outperforms some typical recommended methods.To analyse the privacy preserving for social recommendation,a hierachical network construction method based on hierachical random graph is proposed.This method utilizes hierarchical binary tree model to match the attribute information of input network and the connection probability to substitute the edges between nodes.On this basis,the noise will be injected into the probability set and a sanitized network will be generated to march the attributes of the input network which meet the requirments of -differential privacy.Moreover,a social recommendation utility measurement is given to address the impact of sanitized network on social recommendation.The measurement defines the exchangeability and concentration characteristics of utility function,and analyzes the privacy boundaries of sanitized network by calculating the common neighbor nodes.In summary,this thesis studies the social perceptual model for social recommendation,a random walk mechanism on trust network and privacy preserving mechanism with sanitized network.It gives a unified online social application scenario method to portray the social context and investigates the social tie strength and social trust in different online social media.It proposes a social trust network construction method based on target entity's social network structure and a sanitized network construction method based on the definition of differential privacy.They are of great significance in both theory and practice to promote the reasearch on the privacy preserving issues in social recommendation.